Xiaomi: MiMo-V2-Pro

Processes up to 1 million tokens in a single context window, enabling agents to maintain extended conversation histories, large document sets, and complex multi-step reasoning chains without context truncation. The model architecture supports this through optimized attention mechanisms and memory-efficient transformer implementations, allowing agents to reference prior interactions and accumulated knowledge across extended sessions without losing critical context.

Supports structured function calling and tool invocation within agentic loops, enabling the model to autonomously decide when to call external APIs, execute code, or delegate tasks. The model outputs structured JSON-formatted tool calls that integrate with standard agent frameworks, handling the decision logic for tool selection, parameter binding, and execution sequencing without requiring external routing layers.

Generates, completes, and analyzes code across multiple programming languages with context-aware understanding of syntax, semantics, and best practices. The model leverages its 1T parameter scale and agentic training to produce code that integrates with existing codebases, handle complex refactoring tasks, and provide architectural recommendations based on full codebase context.

Maintains coherent, contextually-aware multi-turn conversations with the ability to reference prior exchanges, correct misunderstandings, and build on previous context. The 1M token window enables the model to preserve full conversation history without summarization, allowing for natural dialogue that spans dozens or hundreds of exchanges while maintaining consistency in tone, knowledge, and reasoning.

Extracts structured information from unstructured text and generates valid JSON outputs conforming to specified schemas. The model uses its reasoning capabilities to parse complex documents, identify relevant entities and relationships, and format outputs according to developer-specified schemas, with support for nested structures, arrays, and type validation.

Synthesizes information across large documents or document sets to produce coherent summaries, identify key insights, and answer questions based on comprehensive document understanding. The 1M token window allows the model to process entire books, research papers, or document collections in a single pass, enabling synthesis without intermediate summarization steps that lose nuance.

Decomposes complex problems into reasoning steps, providing transparent explanations for conclusions and recommendations. The model uses chain-of-thought patterns to work through multi-step logic, mathematical reasoning, and decision-making processes, outputting both final answers and the reasoning path used to arrive at them.

Generates responses that adapt to context, user preferences, and communication style, maintaining consistency in tone, formality, and approach across interactions. The model uses contextual understanding to match communication style to audience (technical vs non-technical, formal vs casual) and adjusts complexity and depth based on inferred user expertise.

Integrates reasoning across text and code domains, enabling the model to explain code in natural language, generate code from descriptions, and reason about code behavior and correctness. The model understands both programming semantics and natural language explanations, enabling bidirectional translation between code and prose.